Regional landslide hazard assessment in the upper Yellow River using an integrated Scoops3D-CNN model

Landslides along large reservoir corridors pose substantial threats to human life, infrastructure, and hydropower facilities, particularly in regions characterized by complex geological conditions and frequent reservoir water-level fluctuations. However, conventional regional landslide hazard assessment methods often have difficulty simultaneously incorporating physically based slope-stability information and characterizing the complex nonlinear relationships among multiple conditioning factors, while the influence of reservoir operation is not always adequately represented. These challenges are exemplified in the Longyangxia-Jishixia section of the upper Yellow River in China, where multiple cascade reservoirs coexist with widespread landslide hazards. To address these limitations, this study proposes an integrated Scoops3D-CNN model for quantitative landslide hazard assessment, developed on the basis of field investigations, remote sensing interpretation, and a comprehensive analysis of landslide-forming conditions. Principal component analysis (PCA) was applied to physical model-derived stability information and statistical conditioning factors, and the resulting components were integrated into the CNN at an optimal contribution ratio of 60% and 40%, respectively, as determined through sensitivity analysis of model performance. In addition to conventional geological, geomorphological, meteorological, and hydrological factors, the maximum rate of reservoir water-level change over a complete fluctuation cycle was incorporated to characterize the potential influence of rapid waterlevel fluctuations on reservoir-bank slope stability. By incorporating Scoops3D-derived slope-stability information into the data-driven framework, the proposed model provides a physically informed representation while retaining the CNN's capability to characterize complex nonlinear relationships among multiple conditioning factors. Model performance was evaluated using an independent test dataset, achieving an AUC of 0.89, higher than those of the standalone Scoops3D model (0.79) and the conventional CNN model (0.87). Spatial comparison with the mapped landslide inventory further showed that the predicted high-hazard zones were consistent with the observed landslide distribution. Overall, the proposed Scoops3D-CNN framework provides a more accurate and reliable approach for regional landslide hazard assessment in reservoir-regulated mountainous areas than traditional single-model methods.